Skip to contentSkip to conversation

tcp: add a global sysctl to control TCP delayed ack

PATCH In review Cong Wang 6 participants 8 messages Apr 4, 2013 – Apr 8, 2013

Subsystems (4)
lore
Share & tools
lore

Cong Wang

author
#0 lore raw

PATCH 1/1 tcp: add a global sysctl to control TCP delayed ack

From: Cong Wang <[email protected]>

Change from RFC:
* make the sysctl per netns

According to previous discussion, it seems there is no reasonable heuristics.

Similar to TCP_QUICK_ACK option, but for people who can't modify the source code and still wants to control TCP delayed ACK behavior.

David, do you still have any objection?

Cc Eric Dumazet <[email protected]>
Cc Rick Jones <[email protected]>
Cc Stephen Hemminger <[email protected]>
Cc "David S. Miller" <[email protected]>
Cc Thomas Graf <[email protected]>
CC David Laight <[email protected]>

---

Documentation/networking/ip-sysctl.txt +5 -0
index f98ca63..9b39681 100644
@@ -572,6 +572,11 @@ tcp_challenge_ack_limit - INTEGER
 	in RFC 5961 (Improving TCP's Robustness to Blind In-Window Attacks)
 	Default: 100
 
+tcp_quick_ack - BOOLEAN
+	Globally enables or disables TCP delayed ACK. The applications
+	can still change the quick ACK mode by TCP_QUICK_ACK option.
+	Default: off
+
 UDP variables:
 
 udp_mem - vector of 3 INTEGERs: min, pressure, max
include/net/netns/ipv4.h +1 -0
index 2ba9de8..f03298a 100644
@@ -63,6 +63,7 @@ struct netns_ipv4 {
 	int sysctl_icmp_errors_use_inbound_ifaddr;
 
 	int sysctl_tcp_ecn;
+	int sysctl_tcp_quick_ack;
 
 	kgid_t sysctl_ping_group_range[2];
 	long sysctl_tcp_mem[3];
net/ipv4/sysctl_net_ipv4.c +9 -0
index fa2f63f..1db1780 100644
@@ -837,6 +837,13 @@ static struct ctl_table ipv4_net_table[] = {
 		.mode		= 0644,
 		.proc_handler	= ipv4_tcp_mem,
 	},
+	{
+		.procname	= "tcp_quick_ack",
+		.data		= &init_net.ipv4.sysctl_tcp_quick_ack,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec,
+	},
 	{ }
 };
 
@@ -866,6 +873,8 @@ static __net_init int ipv4_sysctl_init_net(struct net *net)
 			&net->ipv4.sysctl_ping_group_range;
 		table[7].data =
 			&net->ipv4.sysctl_tcp_ecn;
+		table[9].data =
+			&net->ipv4.sysctl_tcp_quick_ack;
 
 		/* Don't export sysctls to unprivileged users */
 		if (net->user_ns != &init_user_ns)
net/ipv4/tcp_input.c +2 -1
index 6d9ca35..a1b44f3 100644
@@ -3781,7 +3781,8 @@ static void tcp_fin(struct sock *sk)
 	case TCP_ESTABLISHED:
 		/* Move to CLOSE_WAIT */
 		tcp_set_state(sk, TCP_CLOSE_WAIT);
-		inet_csk(sk)->icsk_ack.pingpong = 1;
+		if (!sock_net(sk)->ipv4.sysctl_tcp_quick_ack)
+			inet_csk(sk)->icsk_ack.pingpong = 1;
 		break;
 
 	case TCP_CLOSE_WAIT:
net/ipv4/tcp_output.c +3 -2
index af354c98..130fc99 100644
@@ -169,8 +169,9 @@ static void tcp_event_data_sent(struct tcp_sock *tp,
 	/* If it is a reply for ato after last received
 	 * packet, enter pingpong mode.
 	 */
-	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
-		icsk->icsk_ack.pingpong = 1;
+	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
+	    !sock_net(sk)->ipv4.sysctl_tcp_quick_ack)
+			icsk->icsk_ack.pingpong = 1;
 }
 
 /* Account for an ACK we sent. */
#1 lore raw
Quoted 5 lines · Cong Wang

From: Cong Wang <[email protected]>

Change from RFC:
* make the sysctl per netns

Quoted 8 lines

According to previous discussion, it seems there is no reasonable heuristics.

Similar to TCP_QUICK_ACK option, but for people who can't modify the source code and still wants to control TCP delayed ACK behavior.

David, do you still have any objection?

I totally understand the objections that were given regarding this patch. But for defense of this patch we also provide a knob to disable slow start after idle, which from my point of view is as "evil" as this change.

Btw I think we should make tcp_slow_start_after_idle namespace aware, too.

maintainer
#2 lore raw

From: Hannes Frederic Sowa <[email protected]>
Date: Fri, 5 Apr 2013 00:48:10 +0200

Quoted 4 lines

I totally understand the objections that were given regarding this patch. But for defense of this patch we also provide a knob to disable slow start after idle, which from my point of view is as "evil" as this change.

I completely disagree, slow start after idle is way too aggressively throwing past history away, so turning that off is much safer.

#3 lore raw
Quoted 10 lines · David Miller

From: Hannes Frederic Sowa <[email protected]>
Date: Fri, 5 Apr 2013 00:48:10 +0200

Quoted 4 lines

I totally understand the objections that were given regarding this patch. But for defense of this patch we also provide a knob to disable slow start after idle, which from my point of view is as "evil" as this change.

I completely disagree, slow start after idle is way too aggressively throwing past history away, so turning that off is much safer.

I've been carrying a per-socket way to control delayed ack in my tree for a while, and it is a big performance gain on wifi networks where the network is basically half-duplex. If I recall correctly, it's worth 50+Mbps throughput improvement in some cases. If it matters, I can post more detailed numbers.

There are drawbacks when you set delayed ack too high: The ramp-up time for TCP takes a good bit longer. But, some applications may want to trade slower startup time for better bulk transport, and at moderate delayed-ack values, the ramp up time is not noticeably impaired.

Thanks,
Ben

Signature · 2 lines
Ben Greear <[email protected]>
Candela Technologies Inc  http://www.candelatech.com
#4 lore raw
In reply to Cong Wang
Quoted 8 lines · * Cong Wang | 2013-04-04 18:16:00 [+0800]:

According to previous discussion, it seems there is no reasonable heuristics.

Similar to TCP_QUICK_ACK option, but for people who can't modify the source code and still wants to control TCP delayed ACK behavior.

David, do you still have any objection?

Maybe:

http://kerneltrap.org/mailarchive/linux-netdev/2010/8/23/6283640

I submitted a identical patch 3 years back. The consensus was to implement this on a per-route tunable - not on a global sysctl tunable. OK, consensus is not the correct wording ... ;-)

Hagen

maintainer
#5 lore raw
In reply to Cong Wang

From: Cong Wang <[email protected]>
Date: Thu, 4 Apr 2013 18:16:00 +0800

Quoted 14 lines

From: Cong Wang <[email protected]>

Change from RFC:
* make the sysctl per netns

According to previous discussion, it seems there is no reasonable heuristics.

Similar to TCP_QUICK_ACK option, but for people who can't modify the source code and still wants to control TCP delayed ACK behavior.

David, do you still have any objection?

Quoted 7 lines
Cc Eric Dumazet <[email protected]>
Cc Rick Jones <[email protected]>
Cc Stephen Hemminger <[email protected]>
Cc "David S. Miller" <[email protected]>
Cc Thomas Graf <[email protected]>
CC David Laight <[email protected]>

I'm not applying a patch that adds a global parameter for an attribute which has per-path scope.

author
#6 lore raw
Quoted 3 lines · David Miller

I'm not applying a patch that adds a global parameter for an attribute which has per-path scope.

Ok, I will make it per-route.

Thanks.

#7 lore raw
Quoted 6 lines · Cong Wang
Quoted 3 lines · David Miller

I'm not applying a patch that adds a global parameter for an attribute which has per-path scope.

Ok, I will make it per-route.

Is setting a per-route attribute to cover the backside of an application for which source code is not available in this day and age really all that much easier than hitting that application with an LD_PRELOAD library to make the desired setsockopt() call?

rick jones
having flashbacks to similar conversations in other contexts in the 1990s...